electronic reprint

Acta Crystallographica Section D

BiologicalCrystallography

ISSN 0907-4449

Editors: E. N. Baker and Z. Dauter

FINDMOL: automated identification of macromolecules inelectron-density maps

E. W. McKee, L. D. Kanbi, K. L. Childs, R. W. Grosse-Kunstleve, P. D. Adams, J. C.Sacchettini and T. R. Ioerger

Copyright © International Union of Crystallography

Author(s) of this paper may load this reprint on their own web site provided that this cover page is retained. Republication of this article or itsstorage in electronic databases or the like is not permitted without prior permission in writing from the IUCr.

Acta Cryst. (2005). D61, 1514–1520 McKee et al. � FINDMOL

research papers

1514 doi:10.1107/S0907444905027332 Acta Cryst. (2005). D61, 1514–1520

Acta Crystallographica Section D

BiologicalCrystallography

ISSN 0907-4449

FINDMOL: automated identification ofmacromolecules in electron-density maps

E. W. McKee,a L. D. Kanbi,a

K. L. Childs,a R. W. Grosse-

Kunstleve,b P. D. Adams,b

J. C. Sacchettinic and

T. R. Ioergera*

aDepartment of Computer Science, Texas A&M

University, 301 H. R. Bright Building,

3112 Texas A&M University, College Station,

TX 77843, USA, bLawrence Berkeley National

Laboratory, One Cyclotron Road, Building

64R0121, Berkeley, CA 94720, USA, andcDepartment of Biochemistry and Biophysics,

Texas A&M University, 103 Biochemistry/

Biophysics Building, 2128 TAMU,

College Station, TX 77843, USA

Correspondence e-mail: ioerger@cs.tamu.edu

# 2005 International Union of Crystallography

Printed in Denmark – all rights reserved

Automating the determination of novel macromolecular

structures via X-ray crystallographic methods involves

building a model into an electron-density map. Unfortunately,

the conventional crystallographic asymmetric unit volumes

are usually not well matched to the biological molecular units.

In most cases, the facets of the asymmetric unit cut the

molecules into a number of disconnected fragments, rendering

interpretation by the crystallographer significantly more

difficult. The FINDMOL algorithm is designed to quickly

parse the arrangement of trace points (pseudo-atoms) derived

from a skeletonized electron-density map without requiring

higher level prior information such as sequence information or

number of molecules in the asymmetric unit. The algorithm

was tested with a variety of density-modified maps computed

with medium- to low-resolution data. Typically, the resulting

volume resembles the biological unit. In the remaining cases

the number of disconnected fragments is very small. In all

examples, secondary-structural elements such as �-helices or

�-sheets are easily identifiable in the defragmented arrange-

ment. FINDMOL can greatly assist a crystallographer during

manual model building or in cases where automatic model

building can only build partial models owing to limitations of

the data such as low resolution and/or poor phases.

Received 18 June 2005

Accepted 30 August 2005

1. Introduction and background

A key step in automating the determination of macro-

molecular structures via X-ray crystallographic methods

involves building a model consisting of atomic coordinates

into an electron-density map. Some methods such as ARP/

wARP (Perrakis et al., 1999) and RESOLVE (Terwilliger,

2004) work directly from structure factors, applying symmetry

to generate electron density at any arbitrary point in space.

Other methods such as TEXTAL (Holton et al., 2000),

X-Powerfit (Oldfield, 2003) and MAID (Levitt, 2001) must

build within the boundaries of a given map generated for a

specific region of space by the user. Although the asymmetric

unit (ASU) is commonly used in crystallographic data

processing, it is not optimal for model building. The macro-

molecule can be positioned arbitrarily within the ASU volume

and the facets of the conventional ASU volume (Hahn et al.,

1984) can cut the molecule into a number of disconnected

fragments, with symmetry-related portions appearing on

opposite sides of the map. As a result, when building models

manually crystallographers must often visually identify a

contiguous volume of density which contains the protein

macromolecule and define an appropriate mask around a

electronic reprint

symmetrically unique region of space which may be arbitrarily

shaped and which may cross one or more ASU boundaries1.

The ability to automatically identify a region of electron

density containing the macromolecule and determining its

molecular boundaries without the need for human input/

judgement is critical for more fully automating the structure-

determination process, especially for high-throughput appli-

cations such as PHENIX (Adams et al., 2002, 2004). Although

methods for protein masking and discrimination from solvent

regions (Cowtan & Zhang, 1999; Abrahams & Leslie, 1996;

Brunger et al., 1998; Terwilliger, 2004) have been available for

some time, it has proven challenging to robustly select the core

of a macromolecule and determine the correct packing of

domains/subunits.

In this paper, we present a novel algorithm for automated

identification of a region of electron density containing the

macromolecule. FINDMOL starts similarly to other protein-

masking methods by creating a solvent-flattened map to

enhance density within the protein regions. A skeleton of trace

points (Greer, 1974; Ioerger & Sacchettini, 2002) is then

created inside the high-density contours, representing occu-

pied regions of space. Cluster analysis is used to divide the

trace points into local groups that are either disconnected

from each other or have only minimal contact, e.g. across non-

biological crystal contacts (Carugo & Argos, 1997; Dasgupta et

al., 1997; Valdar & Thornton, 2001). Furthermore, crystallo-

graphic symmetry operations are used to extend clusters

across ASU boundaries, allowing complete macromolecules to

be identified by appending adjacent symmetry copies of trace

points from other parts of the ASU. Finally, the resultant trace

points may be used to generate a mask that can be output for

automated map generation and automated model building.

While the routine usually cannot separate subunits packed

in biologically relevant complexes, which often have more

extensive protein–protein interfaces, tests on a variety of

medium-resolution maps (2.0–3.0 A) show that FINDMOL

can effectively identify regions of electron density containing a

contiguous and unique macromolecule for building. This

facilitates automation of structure determination and can also

benefit crystallographers as a tool for manual model building.

The FINDMOL algorithm has been incorporated into

PHENIX, a comprehensive crystallographic computing plat-

form for structure determination (Adams et al., 2002, 2004).

2. Data preparation

In preparation for running FINDMOL, a density-modified

electron-density map is produced using a method such as DM

(Cowtan & Zhang, 1999) and then scaled and skeletonized

with CAPRA (Ioerger & Sacchettini, 2002). The scaling

routine is designed to normalize the magnitudes of the density

values. The tracing routine returns a set of trace points

(pseudo-atoms), a skeletonized surrogate for the high-density

regions of the map occupied by the macromolecule, derived

from an interpolated 0.5 A grid over the region, which

approximates the medial axis of a 1� contour. In order to

allow these trace points to span the entire ASU, the electron

density is computed over the ASU and a small border. The

initial set of trace points is then pruned by elimination of

points which lie in the border. The identification of trace

points lying entirely in the border is facilitated through use

of the direct_space_asu class in the CCTBX (Grosse-

Kunstleve et al., 2002), the open-source crystallographic

library component of the PHENIX software suite (Adams et

al., 2002, 2004). The pruned set of trace points forms the input

for the FINDMOL algorithm.

3. The FINDMOL algorithm

3.1. Definitions

Given a trace point i, another trace point j is said to be a

neighbor of i if distance(i, j) � ", where distance is the

Euclidean distance and " is a user-defined parameter. The

default value for " is 6 A. Adjusting the neighborhood radius

in a reasonable range (5–9 A) may help prevent trace-point

arrangements bridging symmetry-related molecules in crystal

structures with tightly packed molecules.

The set of all neighbors, the neighborhood, for a point i,

N(i), is defined as N(i) = { j| j is a neighbor of i}. The neigh-

borhood density of i, D(i), is defined to be number of neigh-

bors in the set |N(i)|/(4�"3).

For the purpose of these definitions, a trace point and all its

symmetry-equivalent copies are considered. Essentially, the

FINDMOL algorithm is a means of selecting one member

from each set of symmetry-equivalent trace points so that the

selected points are packed together in a contiguous part of

space.

3.2. Methods

3.2.1. Rationale and overview. Typically, the neighborhood

densities D(i) in the core regions of proteins are higher than

those in surface regions. In the solvent regions, trace points

are very sparse owing to the prior application of density-

modification procedures. A gradual transition from high to

low neighborhood density occurs at the protein surface and

depends on ". Furthermore, in most cases crystal contacts

involve less surface area than biological contacts (Dasgupta et

al., 1997); thus, there will often be a difference in the density of

trace points between the two. This will lead to trace points

near a crystal contact having a smaller neighborhood density

than trace points near a biological contact. Thus, the overall

strategy of the FINDMOL algorithm is to select trace points in

a sequence of decreasing neighborhood density, effectively

building up a region starting from the core of the protein and

expanding outwards to the surface. In such an ordering, it is

expected that biological contacts will be crossed before crystal

contacts are considered (Fig. 1).

As input, FINDMOL takes a set of structure factors and

phases and then reassembles trace points from a skeletonized

research papers

Acta Cryst. (2005). D61, 1514–1520 McKee et al. � FINDMOL 1515

1 A unique and contiguous macromolecule is important for subsequentrefinement steps, to avoid steric clashes by symmetry and to allow propercomputation of geometric and non-bonded constraints, preventing severedistortions of the macromolecule.

electronic reprint

electron-density map generated with data up to a resolution of

2.8 A. High-resolution data sets may generate maps

containing strong side-chain density at crystal contacts,

whereas at 2.8 A resolution side-chain density is typically not

so well defined. As a result of limiting the resolution, we

obtain fewer trace points near the surface of the protein. This

prevents FINDMOL from bridging over crystal contacts.

It was also observed that small clusters of trace points can

be found on the surface of the molecule, representing surface

side chains or ordered solvent. It was found that better results

can be obtained if these points are removed prior to the main

FINDMOL algorithm owing to the reduction in contact

between neighboring molecules in the crystal structure. Such

points typically manifest themselves as small isolated clusters

that are separated from the main cluster of points by a

distance greater than one grid unit (0.5 A) of the electron-

density map. Hence, these peripheral points are removed by

cluster analysis.

The maximum distance between two adjacent grid points,

approximated as a cubic grid, is approximately (0.5� 3)1/2 A =

0.877 A. To locate these points, the asu_clusters object of

the CCTBX is used to tabulate connected components of

points within a 1.0 A radius. This clustering operation can be

visualized by letting the trace points be the vertices of a graph.

Any two vertices i and j are considered connected if

distance(i, j) < 1.0. All points which are members of connected

components of size < 3 are eliminated. As a consequence, this

operation also removes other types of noise which may be

found in the map, including isolated regions of density in

solvent regions.

3.2.2. Algorithm. The neighborhood calculations are based

on the asu_mappings and pair_asu_table classes provided

by the CCTBX. These classes were originally developed for

the efficient computation of non-bonded interactions in

refinement (Grosse-Kunstleve et al., 2004), where similar

algorithms are part of the standard repertoire. The

asu_mappings class uses the space-group symmetry to map

the initial set of trace points into the conventional ASU

volume plus a border of width ! to account for all possible

neigbors of trace points lying near a facet of the ASU volume.

The pair_asu_table class builds the list of trace-point pairs

within radius " using a highly efficient algorithm that assigns

the points stored in the asu_mappings object to cubes with

research papers

1516 McKee et al. � FINDMOL Acta Cryst. (2005). D61, 1514–1520

Figure 2A schematic diagram illustrating the FINDMOL algorithm.

Figure 1A comparison of trace=point density in an ASU of �2u globulin. Thisillustrates four chains with associated 222 symmetry. The transparentsolvent = accessible surface of the tetramer is superimposed over theASU. The trace points are sparser at the crystal contact than at thebiological contact, leading to a difference in neighborhood density. Tightextensive interactions exist between each chain of the tetramer. Allfigures were created with CHIMERA (Pettersen et al., 2004).

electronic reprint

edge length ". Distance calculations are performed only for

points assigned to neighboring cubes, thus reducing the time

complexity of the algorithm from O(N2) to O(N) in the

average case.

The next step in the FINDMOL algorithm (Fig. 2) is to

build clusters by grouping the trace points starting from the

conventional ASU volume. The implementation of this step is

based on the considerations outlined in the previous section.

A queue Q of candidate points is initialized and is empty at the

start. All trace points are marked as unassigned. Each point is

initially associated with the identity symmetry operation. The

trace point with the highest neighborhood density (as

provided by the pair_asu_table object) is selected as the

initial most dense point for the putative core of the molecule

and is enqueued in Q along with its current symmetry

operator.

Subsequently, neighbors are added in order of decreasing

neighborhood density to grow the molecular region by

accretion. Each trace point, Pi, in Q is stored as a pair {xi, �i},

where xi is the original coordinate of Pi and �i is the symmetry

operation required for computation of this symmetric copy of

Pi. At each iteration, the algorithm selects the point Pi with the

maximum neighborhood density from Q. If this point is

already marked as assigned, Pi is discarded and the next

iteration is started. Otherwise, Pi is now marked as assigned

and the following procedures carried out upon it. Coordinates

for Pi are calculated through use of its initial coordinate and its

symmetry operation. All neighbors of this point are then

examined. Each neighbor Pj which is not yet marked as

assigned is enqueued into Q. The required symmetry opera-

tion, �j is computed by

�j ¼ �i��1ia �ja: ð1Þ

In this equation, �ia and �ja are the symmetry operations that

map Pi and Pj, respectively, from the current location into the

ASU volume as tabulated in the pair_asu_table and

asu_mappings objects. Once Q is empty, the algorithm

terminates. In some cases of disconnected density the trace

points devolve into more than one cluster. In this case, not all

points will be marked as assigned after the first pass and the

set of unassigned points is used to rerun the algorithm to

collect additional clusters. This is repeated until all points are

assigned. Each pass generates a cluster rooted in the

conventional ASU volume and growing in an arbitrary

direction. Thus, the final position of the clusters with respect to

one another is arbitrary.

Experiments were also performed to validate that the

heuristic of queuing points based on neighborhood density is

reasonable. As a comparison, a simple FIFO (‘first in first out’)

queue was substituted for the priority queue. In this way, the

algorithm can be envisioned as starting with the point of

highest neighborhood density and expanding in concentric

shells of trace points until a single symmetry copy of each trace

research papers

Acta Cryst. (2005). D61, 1514–1520 McKee et al. � FINDMOL 1517

Figure 3The largest percentage of trace points over the final protein model in anASU for a given neighborhood radius " (5, 6, 7, 8 and 9 A). A key to thePDB codes is given in Table 1.

Table 1Nine experimental electron-density maps were scaled and skeletonized and then defragmented using the FINDMOL algorithm.

Seven data sets (i–vii) were fully defragmented with >95% of the trace points repositioned over the protein macromolecule. Isocitrate lyase and calmodulin (viiiand ix, respectively) were the only proteins that were only partially defragmented; the trace points over the protein macromolecule were �73 and 43%,respectively. These proteins are cases that belong to two types of proteins that currently require additional assembly, either manually or later in the model-buildingprocess.

Protein (PDB code)Resolution(A)

No. of molecules/No. of chainsin the ASU†

Spacegroup

Phasingmethod

No. oftracepoints

% of trace points over finalprotein model or its symmetry-relatedcopy in a neighboring ASU‡

(i) �2u globulin (2a2u) 2.50 1 (4) P212121 MR 4968 99.87 0.10 0.03 —(ii) NADPH-flavin oxidoreductase (1bkj) 2.50 1 (2) P21 SIRAS 3867 99.29 0.52 0.19 —(iii) Cyanase (1dw9) 2.40 1 (10) P1 MAD 14785 99.76 0.19 0.05 —(iv) Armadillo-repeat region from �-catenin (3bct) 2.40 1 (1) C2221 MAD 3698 96.60 2.41 0.94 0.05(v) Neuronal synaptic fusion complex (1sfc)§ 2.40 3 (12) I222 SAD 6077 95.81 3.78 0.18 0.09(vi) Ornithine aminotransferase (1gbn) 2.30 3 (3) P3221 MR 10833 95.34 4.58 0.08 —(vii) HiPIP (1iua) 0.80 1 (1) P212121 MR 417 99.51 0.49 — —(viii) Isocitrate lyase (1f61) 2.00 1 (2)} P6522 MAD 7697 73.25 24.60 2.13 0.02(ix) Calmodulin (1exr) 1.10 1 (1) P1 SAD 1122 43.60 38.69 15.2 1.01

† Values are listed as functional molecules, with the number of chains in parentheses. ‡ The percentage of trace points over one contiguous protein of the trace points inmacromolecule in an ASU. All experimental electron-density maps were generated using a resolution cutoff of 2.8 A. The columns are sorted by the amount of overlap. § Fusioncomplex consists of three distinct proteins. } The biological unit spans over two ASUs.

electronic reprint

point is retained. This was demonstrated using our test set of

electron-density maps to perform worse than the algorithm

employing the priority queue.

4. Results and discussion

Density-modified electron-density maps were computed with

a resolution cutoff of 2.8 A for several experimental data sets

for proteins of various shapes, sizes and oligomeric states.

These maps were scaled and skeletonized using routines from

the CAPRA component of TEXTAL (Ioerger & Sacchettini,

2002). The resulting sets of trace points were then pruned and

passed through the cluster-size filter described previously. A

value of " = 6 A was used as the default neighborhood radius

in the experiments below. The impact of this parameter on the

FINDMOL output was found to be minor when FINDMOL

was run on representative maps

varying " from 5 to 9 A (Fig. 3).

Although the amount of a molecule

found depends on " and no " is

consistently best, the variation is

usually minor (<10%). The output

arrangements were then analyzed by a

scoring routine which generates all

symmetry copies of the known struc-

ture and computes the percentage of

output points which corresponded to

each of the symmetry copies. This

determines the number and size of

molecular fragments generated by the

FINDMOL algorithm. We are

primarily interested in recovering a

symmetry-unique set of points asso-

ciated with a single copy of the mole-

cule, regardless of its molecular

boundary or position in space. Hence,

our metric finds the copy with

maximum coverage. Results from a

representative sample of runs are

summarized in Table 1. These results

are taken from a set of FINDMOL

runs over a collection of 50 phased and

density-modified data sets graciously

provided by Dr Paul Adams.

The nine experimental data sets that

were used to test FINDMOL are (i)

�2u globulin (PDB code 2a2u; Chaud-

huri et al., 1999), (ii) NADPH-flavin

oxidoreductase (PDB code 1bkj;

Tanner et al., 1996), (iii) cyanase (PDB

code 1dw9; Walsh et al., 2000), (iv)

armadillo-repeat region from �-catenin

(PDB code 3bct; Huber et al., 1997), (v)

fusion complex (PDB code 1sfc; Sutton

et al., 1998), (vi) human ornithine

aminotransferase (PDB code 1gbn;

Shah et al., 1997), (vii) HiPIP (PDB

code 1iua; Liu et al., 2002), (viii) iso-

citrate lyase (PDB code 1f61; Sharma et

al., 2000) and (ix) calmodulin (PDB

code 1exr; Wilson & Brunger, 2000).

Over most of the maps (seven out of

nine), FINDMOL is able to locate

large contiguous portions (>95%) of a

single copy of a protein/complex. The

research papers

1518 McKee et al. � FINDMOL Acta Cryst. (2005). D61, 1514–1520

Figure 4Red trace points, skeletonized electron-density map covering a single ASU (red). Green trace points,skeletonized electron density map defragmented by FINDMOL. The rainbow colored ribbons arethe published protein model of (a) NADPH-flavin oxidoreductase, (b) cyanase, (c) neuronal synapticfusion complex and (d) human ornithine aminotransferase. FINDMOL fully defragments the proteinmacromolecule (a–d). All figures were created with CHIMERA (Pettersen et al., 2004).

electronic reprint

two examples where the defragmentation does not fully

succeed are calmodulin and isocitrate lyase (discussed below).

However, even in these cases FINDMOL outputs just two and

three fragments, respectively.

�2u globulin consists of four polypeptide chains of 158

amino acids, with each having a closed eight-stranded �-barrel

structure. The four chains form a tetramer (Chaudhuri et al.,

1999). Almost all of the trace points (>99%) in the ASU are

correctly regrouped over the tetramer by FINDMOL

(Table 1).

NADPH-flavin oxidoreductase consists of two protein

chains of 230 amino acids that form a homodimer in the ASU;

upon dimer formation 4834.6 A2 of solvent-accessible surface

area is lost. NADPH-flavin oxidoreductase reduces FMN to

luciferase, which is important for bacterial bioluminescence

(Tanner et al., 1996). Almost all of the FINDMOL trace points

(>99%) are found over the dimer (Table 1). Tight intraprotein

packing ensured correct defragmentation by FINDMOL.

Cyanase consists of ten polypeptide chains each consisting

of 156 amino acids that assemble to form a dimer of penta-

mers. The result is a decamer that is essential in forming the

active site of the enzyme (Walsh et al., 2000). The core of each

chain consists of four helices forming a folded leaf arrange-

ment with additional subdomains. Almost 100% of the

FINDMOL trace points are found over the protein macro-

molecule owing to the tight protein–protein interactions

between neighboring protein chains in the ASU (Fig. 1a).

Cyanase has an extensive solvent-accessible area of 6345.8 A2

that is lost upon decamer assembly in the ASU (Henrick &

Thornton, 1998). The oligomer is fractured into several

symmetry-related pieces in the conventional ASU.

FINDMOL fully identifies the protein macromolecule from

the electron density extended across several ASUs (Fig. 4b).

Armadillo-repeat region from �-catenin is a 42-amino-acid

sequence motif from murine �-catenin consisting of one

monomer of 457 amino acids. 12 copies of the armadillo repeat

are arranged into two curved layers: a right-handed superhelix

layer and an �/� layer (Huber et al., 1997). The armadillo

repeats have extensive intrachain interactions that enable

FINDMOL to reorganize �97% of the trace points over the

monomer (Table 1). There is little to no biological interaction

between neighboring protein molecules and those that exist

can be assigned as crystal contacts, i.e. surface contacts which

are �400A2 (Dasgupta et al., 1997).

Neuronal synaptic fusion complex is a three-molecule

complex consisting of syntaxin-1A, synaptobrevin-II and

SNAP-25A. Each molecule consists of a four-helix bundle

(Fig. 4e) that forms both a non-globular and unusually twisted

oligomer (Sutton et al., 1998). The fusion complex is involved

in the fusion of vesicles and membranes and its non-globular

nature did not challenge the FINDMOL algorithm. It

correctly defragments this unusual protein with almost 96% of

the trace points being returned over the trimer (Fig. 4c).

Human ornithine aminotransferase consists of three poly-

peptide chains consisting of 402 amino acids and forms a

homodimer. The ASU contains three molecules, two of which

form a homodimer; the third forms a homodimer with itself

across a symmetry axis. The protein has two major domains,

one with a three-layer �/�/� sandwich, while the second

domain is smaller with only a two-layer �/�-sandwich struc-

ture (Shah et al., 1997). FINDMOL correctly repositions

�95% of the trace points over the two domains (Table 1).

HiPIP is a small iron–sulfur protein (83 amino acids),

consisting of a five helices and three �-strands connected

mainly by �-turns (Liu et al., 2002). HiPIP has relatively few

secondary-structural features and as a globular protein

FINDMOL manages to reassemble almost all of the trace

points correctly over it (Table 1), as expected.

Isocitrate lyase from Mycobacterium tuberculosis is a

homotetramer in solution. However, isocitrate lyase crystal-

lizes as a dimer of two pairs of entwined monomers in the

ASU (Sharma et al., 2000). Each monomer consists of 418

amino acids, consisting of a TIM �/�-barrel fold with addi-

tional subdomains that form a domain-swapped dimer. The

biological tetramer is formed from two of these dimers

bisected by a crystallographic twofold axis. FINDMOL only

partially identified the protein molecule owing to tight packing

between the neighboring symmetrically-related dimers that

form the biological tetramer. An extensive solvent-accessible

area of 7127.5 A2 is lost upon tetramer assembly (Henrick &

Thornton, 1998). It is expected that FINDMOL would

experience difficulties with identifying the biological unit

owing to the symmetric redundancy. Isocitrate lyase’s ‘bio-

logical symmetry’ coincides with the crystallographic

symmetry; hence, its biological unit spans over two ASUs.

Only 73.3% of trace points are returned over one dimer and

the rest are over neighboring dimers (Table 1). In cases such as

this, FINDMOL is likely to return protein segments (domains,

folds or secondary structure) which can be visually identified

and moved to the correct position.

Calmodulin is a case of weak intraprotein interactions; a

single disordered helix connects two domains. In this case,

interactions with neighboring molecules are more extensive

than those found within the protein itself. Calmodulin has a

single 146-amino-acid chain consisting of two pairs of EF-hand

units, each having two helices connected by a calcium-binding

loop (Wilson & Brunger, 2000). The two domains are

connected by a long isolated helix, which has relatively weak

density [by D(i) calculation], making a ‘dumb-bell’ shape.

FINDMOL partially defragments the protein macromolecule,

returning approximately 44, 39 and 15% of the trace points

over three distinct domains of calmodulin (Table 1) and these

can be later reassembled during model building. In this case,

FINDMOL essentially returns two domains from separate

symmetry copies, whose interface is denser than the inter-

molecular helical linker.

5. Conclusions

In this paper, we present the FINDMOL algorithm for the

automated identification of macromolecules in medium- to

low-resolution density-modified electron-density maps.

FINDMOL can assist automated model building by defining a

contiguous symmetry-unique region of space for model

research papers

Acta Cryst. (2005). D61, 1514–1520 McKee et al. � FINDMOL 1519electronic reprint

building. The algorithm is based on the cluster analysis of

neighborhood densities of trace points obtained via skeleto-

nization. FINDMOL does not require knowledge of the

number of molecules in the ASU and is also independent of

sequence information. Our results show that the FINDMOL

algorithm is able to identify boundaries of complete mole-

cules/complexes in low-resolution electron-density maps,

regardless of where they are located with respect to the

conventional ASU volume. FINDMOL also performs well for

unusually shaped protein molecules such as the neuronal

synaptic fusion complex and the armadillo-repeat region from

�-catenin. This demonstrates that FINDMOL is not limited to

globular proteins alone. However, the current methodology

has two limitations. Firstly, FINDMOL cannot easily separate

tightly packed protein complexes. Secondly, FINDMOL is

challenged by macromolecules with biological symmetry that

coincides with crystallographic symmetry, e.g. a dimer inter-

face across a twofold axis. Proteins with internal regions of

weak electron density may also be difficult to assemble.

Preliminary experiments indicate that some improvements are

possible based on the ideas of ‘dilation and erosion’ estab-

lished in the field of mathematical morphology (Serra, 1982),

but it is clear that tightly packed complexes can be built

correctly only if the crystallographic symmetry is reconsidered

after advancing the model building to the point of docking the

sequence. However, even in these cases FINDMOL will be

helpful since it is likely to greatly reduce the number of

required rearrangement steps.

6. Availability and performance

The FINDMOL algorithm described in the paper is imple-

mented as part of the TEXTAL suite (Ioerger & Sacchettini,

2003), which is available as a standalone from the corre-

sponding author and also as a component of the PHENIX

suite (Adams et al., 2002) available from http://

phenix-online.org. A map as large as cyanase which contains

ten molecules (1560 amino acids) and 14 785 trace points in

the P1 space group can be scaled, traced and then defrag-

mented by FINDMOL to cover the homodecamer within 90 s

(AMD Opteron 2.6 GHz).

The authors would like to thank all the members of the

TEXTAL group, Tod Romo, Reetal Pai, Kreshna Gopal and

Jacob Smith, for their valuable criticism and support for this

research. This work was supported by NIH grant GM-63210.

We especially thank Peter Zwart for his suggestions and

feedback in the preparation of this manuscript. Crystal

structure data sets were graciously provided to Paul Adams by

the Berkeley Structure Genomics Center, A. Brunger, G.

Gilliland, E. Gordon, O. Herzberg, J. Sacchettini and J. Smith.

References

Abrahams, J. P. & Leslie, A. G. W. (1996). Acta Cryst. D52, 30–42.Adams, P. D., Gopal, K., Grosse-Kunstleve, R. W., Hung, L. W.,

Ioerger, T. R., McCoy, A. J., Moriarty, N. W., Pai, R. K., Read, R. J.,Romo, T. D., Sacchettini, J. C., Sauter, N. K., Storoni, L. C. &Terwilliger, T. C. (2004). J. Synchrotron Rad. 11, 53–55.

Adams, P. D., Grosse-Kunstleve, R. W., Hung, L. W., Ioerger, T. R.,McCoy, A. J., Moriarty, N. W., Read, R. J., Sacchettini, J. C., Sauter,N. K. & Terwilliger, T. C. (2002). Acta Cryst. D58, 1948–1954.

Brunger, A. T., Adams, P. D., Clore, G. M., DeLano, W. L., Gros, P.,Grosse-Kunstleve, R. W., Jiang, J.-S., Kuszewski, J., Nilges, M.,Pannu, N. S., Read, R. J., Rice, L. M., Simonson, T. & Warren, G. L.(1998). Acta Cryst. D54, 905–921.

Carugo, O. & Argos, P. (1997). Proteins, 28, 29–40.Chaudhuri, B. N., Kleywegt, G. J., Bjorkman, J., Lehman-McKeeman,

L. D., Oliver, J. D. & Jones, T. A. (1999). Acta Cryst. D55, 753–762.Cowtan, K. D. & Zhang, K. Y. (1999). Prog. Biophys. Mol. Biol. 72,

245–270.Dasgupta, S., Iyer, G. H., Bryant, S. H., Lawrence, C. E. & Bell, J. A.

(1997). Proteins, 28, 494–514.Greer, J. (1974). J. Mol. Biol. 82, 279–301.Grosse-Kunstleve, R. W., Afonine, P. V. & Adams, P. D. (2004). IUCr

Comput. Commun. Newsl. 3, 19–36. http://www.iucr.org/iucr-top/comm/ccom/newsletters/2004jan/index.html.

Grosse-Kunstleve, R. W., Sauter, N. K., Moriarty, N. W. & Adams, P. D.(2002). J. Appl. Cryst. 35, 126–136.

Hahn, T., Shmueli, U. & Wilson, A. J. C. (1984). International Tablesfor Crystallography. Dordrecht: Kluwer Academic PublishersGroup.

Henrick, K. & Thornton, J. M. (1998). Trends Biochem. Sci. 23, 358–361.

Holton, T., Ioerger, T. R., Christopher, J. A. & Sacchettini, J. C.(2000). Acta Cryst. D56, 722–734.

Huber, A. H., Nelson, W. J. & Weis, W. I. (1997). Cell, 90, 871–882.Ioerger, T. R. & Sacchettini, J. C. (2002). Acta Cryst. D58, 2043–2054.Ioerger, T. R. & Sacchettini, J. C. (2003). Methods Enzymol. 374, 244–

270.Levitt, D. G. (2001). Acta Cryst. D57, 1013–1019.Liu, L., Nogi, T., Kobayashi, M., Nozawa, T. & Miki, K. (2002). Acta

Cryst. D58, 1085–1091.Perrakis, A., Morris, R. & Lamzin, V. S. (1999). Nature Struct. Biol. 6,

458–463.Pettersen, E. F., Goddard, T. D., Huang, C. C., Couch, G. S.,

Greenblatt, D. M., Meng, E. C. & Ferrin, T. E. (2004). J. Comput.Chem. 25, 1605–1612.

Oldfield, T. (2003). Methods Enzymol. 374, 271–300.Serra, J. P. (1982). Image Analysis and Mathematical Morphology.

New York: Academic Press.Shah, S. A., Shen, B. W. & Brunger, A. T. (1997). Structure, 5, 1067–

1075.Sharma, V., Sharma, S., Hoener zu Bentrup, K., McKinney, J. D.,

Russell, D. G., Jacobs, W. R. Jr & Sacchettini, J. C. (2000). NatureStruct. Biol. 7, 663–668.

Sutton, R. B., Fasshauer, D., Jahn, R. & Brunger, A. T. (1998). Nature(London), 395, 347–353.

Tanner, J. J., Lei, B., Tu, S. C. & Krause, K. L. (1996). Biochemistry, 35,13531–13539.

Terwilliger, T. (2004). J. Synchrotron Rad. 11, 49–52.Valdar, W. S. & Thornton, J. M. (2001). Proteins, 42, 108–124.Walsh, M. A., Otwinowski, Z., Perrakis, A., Anderson, P. M. &

Joachimiak, A. (2000). Structure Fold. Des. 8, 505–514.Wilson, M. A. & Brunger, A. T. (2000). J. Mol. Biol. 301, 1237–

1256.

research papers

1520 McKee et al. � FINDMOL Acta Cryst. (2005). D61, 1514–1520

electronic reprint